Home › Story

We let an AI issue 894 crypto signals — the honest scorecard

Updated 2026-08-28 · 894 signals tracked

• 60.7%Win rate
• 1.27Avg R
• 50TP2 hits
• 894Signals

The setup

Every ~15 minutes, an AI model (Anthropic Claude) reads dozens of live features per asset — momentum, order flow, funding, open interest, volatility, macro — and decides whether there's a risk-defined trade worth taking on Bitcoin, Ethereum and other perpetual futures. Each call is a clean, testable trade: an entry, a stop-loss and two take-profit targets — TP1, the win line, and TP2, the stretch.

The honest scorecard

Across 894 published signals: 60.7% win rate at 1.27 average reward-to-risk — 455 wins against 295 losses, with expiries excluded and 50 signals going on to hit TP2. Not every trade wins — and crucially, the losers are all still on the record. A positive average reward-to-risk above 1 with a win rate near 50% is a genuinely positive-expectancy result, and it is fully auditable in the track record.

What we learned

• Transparency is rare and valuable — publishing losses builds more trust than hiding them.
• "Accuracy" alone is misleading; reward-to-risk matters as much as win rate.
• An AI can apply a disciplined process consistently, but no model wins every trade.

See for yourself

Watch the live signals on the dashboard, download the raw CSV, or read how accurate are AI trading signals.

How the pipeline actually decides

Every signal starts with 144 engineered features, rebuilt every 15 minutes from 13 live sources: exchange books from Delta Exchange, Binance, Bybit, Deribit and Hyperliquid; market and derivatives data from CoinGlass, Coinalyze and CoinGecko; and macro feeds from DefiLlama, Alternative.me, Farside and yfinance. Momentum, funding rate, open interest and volatility all land in the same call. Claude Opus 4.8 reads that feature set at temperature zero. Forced tool-calling keeps the output strict JSON, never free-form prose. A code-level risk gate then checks reward-to-risk, rejecting anything under 1.0 and clamping every target to a 1.0-1.5 band. Nothing skips that gate on the way to the engineering pipeline's publisher. The structure stays fixed whether the market is calm or violent: no discretionary override, no edits after the fact.

A signal, from generation to close

A signal doesn't finish the moment it publishes. The tracker checks every open position each cycle against three exits: TP1 first, TP2 beyond it, and the stop-loss below entry. Touch TP1 or TP2 first and the trade counts as a win. Touch the stop first and it counts as a loss. Of 894 published signals, 750 were decisive, closed as a clean win or loss inside the tracking window. The remaining 135 expired unresolved after 24 hours and never enter the win-rate math. That is why the win-rate denominator is smaller than the total signal count. The model also files a self-reported outcome probability with every signal. That number gets graded after the close with a multiclass Brier score, a separate check on whether stated confidence matched reality, not on whether the trade made money.

Does direction change the odds

Split the record by direction and the two sides don't look identical. Long signals land a 64.3% win rate across 557 calls; short signals land 50.3% across 193. Crypto perpetuals spend more time in uptrends than downtrends over most multi-year windows, and funding costs bite longs and shorts differently depending on the regime. Neither number is fixed. Both move as the 22 July 2026 sample grows and as market structure shifts. Treat the split as a diagnostic, not a rule: it says more about the period covered so far than about which direction to prefer next.

How much should you trust 60.7%

A single win-rate number hides how much a small sample can wobble. The 95% confidence interval around the headline 60.7% currently runs from 57.1% to 64.1%. That range will narrow as 894 grows and widen after any losing streak. A rate that sits comfortably above 50% with a tight interval means something different from the same rate on a handful of trades. Readers who want the statistical case for why sample size matters this much can see the full breakdown at the win-rate benchmark page. The short version: judge this system on hundreds of decisive trades, not on any single week.

What this system does not do

Honesty cuts both ways. The record above is real, but it isn't the whole story of what it takes to trade it.

• It does not account for exchange fees, funding payments or slippage between the published price and your actual fill.
• It does not size positions, manage leverage or tell you how much capital to risk on any one trade.
• It cannot see a regime change coming any better than the data already reflects, and a sharp break from recent conditions can hurt it same as any model.
• It does not adapt mid-trade. Once a signal is published, its entry, stop and targets are fixed until close.
• A positive track record so far is not a guarantee of a positive record next month. Markets change, and the 894-signal sample is still young, running since 22 July 2026.

Trading this from India

Drishti Pro's signals target Delta Exchange India perpetual futures, and the tax rules there are unforgiving. Gains on virtual digital assets are taxed flat at 30% plus cess under Section 115BBH, with no set-off against losses and no carry-forward to future years. A 1% TDS applies to most transactions under Section 194S, and every gain has to be declared in Schedule VDA of your income tax return. Only exchanges registered with FIU-IND can legally operate in India, and crypto itself still holds no legal-tender status. None of this changes whether a signal wins or loses. It changes what you keep after it does, and it's worth knowing before the first trade, not after.

See Drishti Pro's live signals Get the app

Educational & illustrative only — not investment advice. Drishti Pro publishes AI-generated trade ideas and their public track record for information. Crypto is volatile and you can lose money. Nothing here is a recommendation to buy or sell any asset. Do your own research.